Hi Product Hunt, I'm one of the founders of Belay. The problem we kept hitting: every new coding agent session starts from zero. The agent said "done" when tests had not run. We corrected the same scoping mistake three days in a row. We fixed it, and the next session made it again, because nothing carried forward. Belay is our answer. It reads the Claude Code and Codex history already on your disk and finds the patterns: the same correction three times, the same false "done", the same wasted detour. From those it proposes lessons, each one traced to the exact turns as evidence. You read them, approve the ones you agree with, narrow or edit the ones you half agree with, reject the rest. Nothing is active until you say so. Then the loop closes. At the start of your next session, your agent is shown the approved lessons as a Mission Pack, and it has to ask you before using them. Belay never edits your project files. If a lesson would change something like CLAUDE.md or AGENTS.md, you get a diff to apply yourself. Two things we were strict about. Everything stays on your machine: secret-scrubbed, encrypted locally, loopback only, no account, no hosted model calls, MIT engine. And we do not claim the lessons make you better yet. We have not measured that, and we will not say it until we have. It is an unsigned alpha for Apple Silicon Macs and Windows 10/11 x64, so expect a Gatekeeper or SmartScreen prompt on first run. macOS: curl -fsSL https://getbelay.vercel.app/install | bash Windows (PowerShell): irm https://getbelay.vercel.app/inst... | iex What lessons would you want your agent to stop relearning? And what would you never let a tool read?

Belay

Lessons from your coding agent's mistakes, carried forward

Lessons from your coding agent's mistakes, carried forward